Research Article
Predictors of High School Students’ Intentions and Behaviors in Using AI Learning Tools: An Extended Theory of Planned Behavior Approach
Synthesis: Zhang and An (2026) fitted an extended theory of planned behavior model to 513 students in three Hangzhou public high schools, asking what predicts intention to use AI learning tools and three behaviors: seeking AI help, evaluating responses, and applying output. Instrumental attitude led the intention paths, ahead of subjective norms, perceived behavioral control, and affective attitude; the model explained 57.3% of intention variance. AI anxiety was unrelated to intention (β = 0.056, p = 0.123) yet negatively predicted all three behaviors, most strongly applying output (β = −0.247). The authors read that split as anxiety acting at enactment rather than during intention formation.
Key Findings
- Usefulness, not enjoyment, carried intention. Students who saw AI tools as ways to understand material, check their own gaps, and save time were more willing to use them than students who found the tools merely pleasant.
- Adult approval outweighed capability. What teachers, parents, and peers expected mattered more than students' confidence in their own skill, so the social message may move intention more than another lesson on tool mechanics.
- Anxiety left intention intact but shrank actual use. Anxious students still said they planned to use AI, yet reported less asking for help, less checking of answers, and less use of AI output.
- The penalty was largest where the stakes are highest. Anxiety's strongest negative path ran to applying AI output, where AI text enters a student's submitted work.
- Confidence predicts who starts, not who judges well. Perceived control mattered most for seeking help and least for evaluating responses, so skill-building buys entry more than judgment.
- Intention tracked consequence, not effort. The intention-behavior link was strongest for applying output, then evaluating, then asking, a gradient from low-stakes questioning to high-stakes adoption.
- Students were moderately positive and only mildly anxious. Usefulness ratings were the highest of the antecedents and anxiety the lowest; checking AI answers was more common than applying them.
A Sample Already Living With These Tools
The study retained 513 valid questionnaires from 550 distributed (93.3% effective response rate), collected over two weeks from Grades 10 and 11 (260 and 253 students aged 15 to 17). Grade 12 was excluded, and all three schools already used tools such as Doubao, ERNIE Bot, and SparkDesk, so the findings fit students already using AI.
Nine constructs used 5-point scales after translation and expert review. Reliability held (Cronbach's alpha 0.81 to 0.90), and the structural model showed good fit (CFI = 0.996, RMSEA = 0.042, SRMR = 0.052).
Intention Rests on Usefulness and Approval
Students who believed AI tools would clarify confusions, reveal their own strengths and weaknesses, and save time were far more willing than students who merely found the tools pleasant. Instrumental attitude was the strongest path (β = 0.447) against a weaker affective attitude path (β = 0.182), and the pattern echoes the technology acceptance literature in a population whose study time runs on assignments and exams.
The strength of subjective norms carries a different message: adult expectations (β = 0.374) mattered more than perceived behavioral control (β = 0.237), so a consistent school-and-home line on legitimate use may move intention more than another skills workshop. Perceived control is not just access: it includes confidence in asking effective questions, interpreting answers, and handling inaccuracies.
Three Behaviors, Three Different Levers
The three behaviors correlated between 0.43 and 0.47, and different predictors dominated each. Perceived control governed entry: students who felt confident asked more often. Evaluation and application leaned more on intention, and the model explained 43.0% of variance in seeking help but 38.5% in the other two.
The gradient is one of difficulty: asking is cheap and reversible, judging takes work but stays private, and incorporating output raises accuracy, originality, and academic integrity questions. The closer behavior comes to graded work, the more it depends on a deliberate decision rather than opportunity.
Anxiety Acts at Enactment, Not at Intention
The distinctive result is a dissociation: anxiety did not shift stated willingness (β = 0.056, p = 0.123), and its indirect effects through intention were nonsignificant (0.013 and 0.014, intervals crossing zero). Its influence ran directly and negatively to behavior: seeking (β = −0.112, p = 0.004), evaluating (β = −0.131), applying (β = −0.247).
The authors' interpretation is dual-route: intention reports draw on reflective judgments about usefulness, approval, and capability, while anxiety surfaces at use, when a student weighs whether an answer is trustworthy and whether a teacher would read reliance as cheating. The weight of the two routes still needs replication. The result also supports a point about student-AI interaction: what students say about AI and what they do with it need separate measurement.
What this means for practice
- Instructors: Skip the novelty pitch. Show what AI does for comprehension, error checking, and revision time; that was the strongest intention path (β = 0.447).
- Administrators: Align the message across teachers, parents, and peers. Subjective norms (β = 0.374) outperformed students' capability, so contradictory rules undercut adoption more than a missing workshop.
- Designers: Target the three stages separately: onboarding builds confidence to start, prompts support verification, and adoption support addresses reliability and disclosure, where anxiety bites hardest (β = −0.247).
- Instructors and counselors: Ask about worry, not just willingness. A student who plans to use AI may still avoid it, and work on misinformation, overreliance, and integrity reaches behavior attitude change will not.
Limitations
- The cross-sectional design supports associations only; reciprocal influence and omitted variables remain possible.
- All constructs, including the three behaviors, are student self-reports at one time point, exposed to social desirability and retrospective bias.
- No learning outcomes were modeled, so the study cannot say whether seeking, evaluating, or applying AI improved understanding.
- Intention was one construct while behavior was split into three, so a matched intention-behavior model went untested.
- Three urban Hangzhou schools already using AI tools limit generalization to rural, private, or less AI-supported settings.
Citation
Zhang, Z., & An, F. (2026). Predictors of High School Students’ Intentions and Behaviors in Using AI Learning Tools: An Extended Theory of Planned Behavior Approach. Behavioral Sciences, 16, 736.